Yager-Rybalov Triple Π Operator as a Means of Reducing the Number of Generated Clusters in Unsupervised Anuran Vocalization Recognition

摘要

The Learning Algorithm for Multivariate Data Analysis (LAMDA) is an unsupervised fuzzy-based classification methodology. The operating principle of LAMDA is based on finding the datum-cluster relationship obtained by means of the Global Adequacy Degrees (GADs) of the Marginal Adequacy Degrees (MADs) of all the data attributes. In comparison with other unsupervised clustering algorithms, LAMDA does not require the number of classes as input parameter; however, in some applications, the quantity of obtained clusters does not correspond with the number of desired classes. Typically, this issue is overcome by merging interrelated clusters within the same class; nevertheless, in some applications the number of generated clusters related to the same class reaches a non-desired and impractical number. In LAMDA, the number of generated clusters is controlled by using a linear mixed connective with an exigency index a. This connective is an unnatural aggregation operator of the MADs, which adds an additional parameter to set up. In this paper, a full reinforcement operator (Yager-Rybalov Triple Ⅱ) is used as aggregation operator for merging the information contained in the MADs. This approach significantly reduces the number of generated classes and suppresses the LAMDA dependence of the parameter a. The proposed approach was tested in a case study related to unsupervised anuran vocalization recognition. A database of advertisement calls of six anuran (frog) species for testing this proposal was selected. All 102 vocalizations were correctly identified (100% of accuracy) and solely the desired classes were generated by the algorithm (establishing a cluster-class bijection).